Continuous Time Warping for Low Bit-Rate Celp Coding

1. An apparatus for modifying a speech signal prior to coding the speech signal, the apparatus comprising: a linear prediction coefficient extraction module for receiving a digital speech signal and for producing a set of linear prediction coefficients; an inverse synthesis filter for receiving the digital speech signal and the linear prediction coefficients and for producing a residual signal; a residual modification module for time warping the residual signal based on a continuous linear warp contour selected from the set of all possible warp contours by calculating only a subset of all possible warp contours, whereby a modified residual signal having a regular rather than irregular array of dominant peaks is produced, wherein no portion of any section of the residual is omitted or repeated in the modified residual; and a synthesis filter for receiving the modified residual signal and the linear prediction coefficients for producing a modified digital speech signal suitable for low bit rate coding.

2. The apparatus according to claim 1 , further comprising a codebook excited linear prediction coder module for receiving the modified digital speech signal and for producing a compressed speech signal.

3. The apparatus according to claim 1 , wherein the residual produced by the inverse synthesis filter comprises irregularly spaced dominant peaks and wherein the residual modification module divides the residual into a series of contiguous, non-overlapping sections, each section containing not more than one dominant peak, derives an idealized signal having a series of regularly-spaced dominant peaks located in a series of sequential sections, associates each section of the residual with a corresponding section of the idealized signal, calculates a linear continuous warp contour for each residual section based on a subset of possible last sample lag values for each residual section within a subrange of possible last sample lag values for each residual section, and modifies the residual by applying the calculated warp contour to the sections of the residual so that any dominant peak in each section of the residual aligns with the dominant peak in the corresponding section of the idealized signal.

4. The apparatus according to claim 3 , wherein the speech signal comprises at least one frame and wherein dividing the residual into a series of contiguous, nonoverlapping sections further comprises analyzing the frame to identify an integer pitch period.

5. The apparatus according to claim 4 , wherein the residual modification module employs co-relation analysis in the open loop in analyzing the frame to identify an integer pitch period.

6. The apparatus according to claim 3 , wherein the residual modification module calculates the subset of all possible warp contours for each residual section by establishing a first sample lag for the first sample of the residual section, identifying a set of candidates for the last sample lag for the last sample of the residual section, the set of candidates consisting of a subset of all possible last sample lag values within a sub-range of all possible last sample lag values, performing a linear interpolation between the first and last samples of the residual section for each candidate last sample lag to create a set of candidate lag contours, applying each candidate lag contour to the residual section to obtain a set of candidate modified residuals, calculating a correlation strength between each candidate modified residual and the corresponding section of the idealized signal to create a set of correlation strengths, deriving an optimal last sample lag for the residual section based on the set of correlation strengths, and deriving a linear continuous warp contour by interpolating linearly over the section from the first sample lag to the derived optimal last sample lag for the residual section.

7. The apparatus according to claim 6 , wherein, in deriving an optimal last sample lag for the residual section based on the set of correlation strengths, the residual modification module segregates the set of correlation strengths into overlapping subsections as a function of the last sample lags used to derive the strengths, represents each subsection as a curve, calculates the maximum value of each curve, wherein the maximum value is selectable from the group consisting of all possible lag values within a range of possible lag values that includes the last sample lags used to derive the strengths in the subsection, and calculates the maximum correlation strength for the section based on the maximum values for the curves of the subsections.

8. The apparatus according to claim 7 , wherein the curve representing each subsection is a polynomial.

9. The apparatus according to claim 8 , wherein the polynomial is a quadratic function.

10. The apparatus according to claim 8 , wherein the subrange of possible last sample lag values for each residual section is selected such that the greatest cumulative shift for any sample in the section upon application of the calculated warp contour will be less than four sample positions.

11. The apparatus according to claim 1 , wherein the speech signal comprises at least one frame and wherein the inverse synthesis filter extracts linear prediction coefficients for the frame, interpolates the linear prediction coefficients for the frame to create linear prediction coefficients for a plurality of sub-frames of the frame, and produces a prediction residual for each sub-frame, whereby the residual signal comprises a set of subframe prediction residuals.

12. A method for using a computer to modify a speech signal prior to coding the speech signal, the method comprising: converting speech to a digital electrical signal; deriving a linear prediction residual signal having irregularly spaced dominant peaks from the digital electrical signal; dividing the residual signal into a series of non-overlapping sections, each section containing a dominant peak; and continuously time warping one or more of the non-overlapping sections such that the modified residual signal comprises regularly spaced dominant peaks and all sections of the residual signal are substantially retained without repetition in the modified residual signal.

13. The method according to claim 12 , wherein the step of continuously time warping one or more of the non-overlapping sections comprises: deriving an idealized signal having a series of regularly-spaced dominant peaks located in a series of sequential sections; and associating each section of the residual signal with a corresponding section of the idealized signal.

14. The method according to claim 13 , wherein the step of continuously time warping one or more of the non-overlapping sections further comprises calculating a linear continuous warp contour for one or more residual sections.

15. The method according to claim 14 , wherein the linear continuous warp contour for the one or more residual sections is based on a subset of possible last sample lag values for each such section within a subrange of possible last sample lag values for each such section.

16. The method according to claim 15 , wherein the step of continuously time warping one or more of the non-overlapping sections further comprises modifying the residual by applying the calculated warp contour to the one or more sections so that any dominant peak in each section of the residual signal aligns with the dominant peak in the corresponding section of the idealized signal, whereby dominant peaks of the modified residual signal are regularly spaced.

17. A computer-readable medium having thereon computer-readable instruction for performing a method of transforming a linear prediction residual signal having irregularly spaced dominant peaks derived from an audio signal to improve the compressibility of the residual signal, the instructions comprising instructions for: dividing the residual signal into a series of non-overlapping sections, each section containing a dominant peak, and continuously time warping one or more of the non-overlapping sections such that the modified residual signal comprises regularly spaced dominant peaks and all sections of the residual signal are substantially retained without repetition in the modified residual signal.

18. The computer-readable medium according to claim 17 , wherein the instructions for continuously time warping one or more of the non-overlapping sections comprise instructions for: deriving an idealized signal having a series of regularly-spaced dominant peaks located in a series of sequential sections; and associating each section of the residual signal with a corresponding section of the idealized signal.

19. The computer-readable medium according to claim 18 , wherein the instructions for continuously time warping one or more of the non-overlapping sections further comprise instructions for calculating a linear continuous warp contour for one or more residual sections.

20. The computer-readable medium according to claim 19 , wherein the instructions for continuously time warping one or more of the non-overlapping sections further comprise instructions for modifying the residual by applying the calculated warp contour to the one or more sections so that any dominant peak in each section of the residual signal aligns with the dominant peak in the corresponding section of the idealized signal, whereby dominant peaks of the modified residual signal are regularly spaced.